@philwhln@AIWarper So this is it working on a face. This can be pushed further and made more accurate. But it's simply triangulation of points over different angles.
@philwhln@AIWarper RAFT takes the front pixels as well and the generated performances, places the pixels in depth based off the front, & compares it for each. Does a measurement and if it's close it keeps that pixel. But I haven't found a solve passed 45degrees yet.
@philwhln@AIWarper Specifically for a face, (but i am working on trying to get it to work for the rest of the body) I am triangulating 51 points across the face, where a ray crosses that is how it's judging where it's position should be (based on the front view). The rest was using RAFT.
@coloroutofplace@AIWarper I think that is very likely to be honest. I went down this direction mainly to see what is possible with a 4d splat at this time since that was what the video was making it seem like.
But yes a 3d track and additional reference images of the other sides of the room would work
@sekiriously@aimikoda I just posted a workflow that technically works with both ltx and minimax h3. H3 did decent but you had to prompt harder to get it to keep the performance locked in. Ltx did a good job but seedance by far was the best.
@AIWarper Haha, it felt like I was on the hunt. I spent way too much time looking at your post. I may have lost my mind grapes a little. Great work by the way.
Definitely. Thatโs a part of what Iโm working into the slip bot. Slowly adding lists of what each model can do well and where it should spend its money. So if it thinks minimax can do a shot it will do that. Still working that out because at the moment it sometimes will choose minimax for something that really needs to be seedance. But Iโm getting it there.
We invented the hybrid pipeline everyone is now pretending they just discovered.Live-action office. AI parrots. A full-size bear walking the aisles.None of the animals were on set.Weโve been doing this for most of the year. Hereโs exactly how:
@icreatelife That would be awesome. Iโve been playing around with stacking agents help speed up the creative process. But also mixing and matching different llm models for different steps based on what they are good at.
@priestessofdada Been refining my individual agents to see how much they can do with one prompt.
Not just shit generation, but edits, sound beds, and post processing effects. Itโs still a slop bot. But itโs getting better as I refine each departments agent.
At our studio we already had a lot of agents helping our artists behind the scenes. We decided to see what would happen if we created a few additional agents and let the director agent go completely autonomous. Full list of models used at the end of the video. Prompt below...
@Sanwerrs0i @theisaacmed@GetContentForge It depends entirely on how you are running the agent / agents. And the range can vary drastically based on how you are handling the memory for these tasks.
Asking opus 5.5 to tell us who it is and what its existence is likeโฆ as far as it understands it. And it can do so in any way it chooses. This is their story.
At our studio we already had a lot of agents helping our artists behind the scenes. We decided to see what would happen if we created a few additional agents and let the director agent go completely autonomous. Full list of models used at the end of the video. Prompt below...